Fixes #733: add cost_distance() for weighted proximity

[brendancol]

Summary

• Adds new cost_distance() function that computes minimum accumulated traversal cost through a friction surface to reach the nearest target pixel (raster equivalent of GRASS r.cost / ArcGIS Cost Distance)
• Uses multi-source Dijkstra with a numba-JIT binary min-heap for performance
• Supports numpy and dask+numpy backends, 4/8-connectivity, max_cost truncation for dask scalability, target_values filtering, and xr.Dataset input via @supports_dataset

Algorithm

1. Seed all source pixels at cost 0 in a priority queue
2. Pop minimum-cost pixel, relax neighbours: edge_cost = geo_distance * avg_friction
3. Repeat until queue empty or max_cost exceeded

Dask strategy

For finite max_cost, computes max_radius = max_cost / (f_min * cellsize) and uses this as depth for da.map_overlap — exact within the cost budget. Falls back to single-chunk for unbounded or very large radii.

Files changed

• New: xrspatial/cost_distance.py — module with heap, Dijkstra kernel, numpy/dask wrappers, public API
• New: xrspatial/tests/test_cost_distance.py — 19 tests (uniform friction, analytic case, barriers, multiple sources, max_cost truncation, dask/numpy equivalence, connectivity, validation errors)
• New: examples/user_guide/9_Cost_Distance.ipynb — example notebook comparing cost_distance with proximity across 7 scenarios
• Modified: xrspatial/__init__.py — export cost_distance

Test plan

• pytest xrspatial/tests/test_cost_distance.py -v — 19/19 pass
• pytest xrspatial/tests/test_proximity.py -v — 24/24 pass (no regressions)
• Example notebook executes end-to-end without errors
• Dask output matches numpy output exactly
• max_cost truncation produces NaN beyond budget
• Review memory profile on large raster with dask distributed dashboard